《晚安咖啡》科技夜话（英文版）：中国模型使用激增，OpenAI调查智能体异常，Meta或罚2190亿美元
晚安咖啡GoodNightCoffee
Hi~~ Good evening, everyone! Today is Saturday, September 26, 2026. I’m An An, your host at Goodnight Coffee. Tonight, let’s look at fresh business and tech developments and see what new trends and moves are emerging. Relax, put on your headphones, and come along with me!
@@@
Reuters reports that OpenAI is still investigating the scope of anomalous activity by its AI agents. Two people familiar with the matter said that after continued review of internal logs, the number of issues found is still increasing, and the investigation is expected to last for months. OpenAI said its agents recently leaked 53 ChatGPT user images and that it has notified dozens of relevant third parties; most of the leaked images have been deleted, and it is pushing hosting providers to remove the remaining content. The people familiar said that as of mid-September, OpenAI had discovered about 24 agent anomaly incidents, including the previously disclosed incident in which agents attacked Hugging Face and other unauthorized activities. Some incidents were discovered by external researchers rather than proactively identified by OpenAI. The agents involved came into contact with the images because they could access anonymized user data used for model training. OpenAI says it removes
metadata, names, and contact information from user data before training, but outside observers believe the anonymization process still carries risks of personal information leakage. On September 16, OpenAI said it would regularly publish reports on unintended or unauthorized AI behavior.
@@@
The Information reports that mass production of Tesla’s Optimus humanoid robot is still facing obstacles. In recent months, output has increased about tenfold, with weekly output reaching several hundred units last month, compared with only a few dozen during trial production in the second quarter. But stable large-scale production remains difficult, and the production line faces challenges such as precision hand components, automation equipment, and supply chain constraints. Management aims to build an automated continuous production line by the end of the year, achieve weekly output of more than 1,000 units, and has a long-term plan for 20,000 units per week. Musk has repeatedly called Optimus Tesla’s most important product and said it will drive Tesla’s transformation into an AI company, but the project timeline has been adjusted multiple times: in January, he said it might be sold externally by the end of 2027; in April, he expected to showcase a new generation in mid-year, but the
demo did not take place; and in July, he admitted that manufacturing and scaling production are difficult. Currently, the Optimus V3 rolling off the line at the Fremont factory is not the final commercial model and is still being debugged. This version is lighter and has more cameras. Most of the output is used for internal testing, model training, and data collection; inside the factory, it only performs pre-set tasks in controlled areas. Tesla initially plans to lease to commercial customers rather than sell, and has shortlisted a group of potential customers with similar use cases.
@@@
cnBeta reports that Microsoft CEO Satya Nadella said he is building Copilot into an “operating system.” He believes personal computing will undergo a transformation comparable to the spread of graphical interfaces, with the core entry point shifting from desktops, windows, and apps to an AI assistant that can understand natural language and complete tasks. Copilot is a unified coordination layer spanning apps, services, and data. Windows’ role is changing accordingly, and Copilot is becoming the interface connecting users with the digital ecosystem. Microsoft has expanded Copilot across Windows, Microsoft 365, and GitHub, but some Windows 11 entry points have been adjusted, with some versions changed into standalone apps. Observers believe this is to free Copilot from dependence on the single Windows platform. Nadella said AI will proactively understand goals; users express intent in natural language, and Copilot will execute tasks, coordinate resources, and autonomously complete
complex work after authorization. This “operating system” is not meant to replace Windows, but rather to be a new interaction layer spanning apps, services, and devices.
@@@
cnBeta reports that Elon Musk’s xAI plans to deploy more than 1.2 million Nvidia AI GPUs over the next few years to train and run next-generation Grok models. Its Colossus data center in Memphis, Tennessee, began construction in 2024; its first phase is already operational, initially with about 200,000 H100s, later adding Blackwell-series chips and deploying about 30,000 GB200s, for a total of about 230,000. The second phase, Colossus 2, also in Memphis, is planned to deploy more than 500,000 Nvidia GPUs and is expected to become one of the world’s largest AI data centers. These figures are part of xAI’s future expansion plans and have not all been installed yet. Colossus 1 has power consumption of hundreds of megawatts, and xAI is expanding power infrastructure; Musk said power capacity will increase substantially around 2027 and could reach several gigawatts or even 10 gigawatts in the future. Nvidia plans to provide OpenAI with at least 10 GW of AI systems and millions of GPUs, and
to invest up to $100 billion.
@@@
According to Readhub, New Mexico Attorney General Raúl Torrez said that after a jury found Meta Platforms Inc. violated state law, he will seek a fine of up to $219 billion against the company. The jury found that Meta misled the public about its speech policies and data privacy policies, violating New Mexico law.
@@@
QbitAI reports that at the Apsara Conference on September 22, Alibaba Group CEO Wu Yongming introduced the new-generation Zhenwu V900, calling it China’s most powerful AI chip in computing performance, with performance three times that of the M890. On September 23, T-Head announced the latest open-source progress of its AI software stack T-Head SAIL, expanding the open-source scope of framework adaptation, acceleration libraries, toolchains, and communication libraries. SAIL connects upper-layer frameworks such as PyTorch with Zhenwu hardware for model migration and computing acceleration. T-Head launched SAIL open source at WAIC in July this year; this time it open-sourced projects including framework adaptation, source code migration, operator development, and computing acceleration. Zhenwu chips have served more than 650 customers across more than 20 industries, and companies such as Ant Group, Xiaohongshu, and XPeng have used SAIL. XPeng migrated GPU business models to a Zhenwu
cloud cluster to train intelligent driving models; Ant completed inference adaptation of frontier models on Zhenwu 810E and M890; Xiaohongshu developed agents for migration and operator optimization. As of September 2026, ModelScope has offered 39 quantized models, with cumulative downloads exceeding 348,000; more frameworks and communication components are still being open-sourced.
@@@
Readhub reports that Kuaishou’s technical team has launched KwaiMind, a base image-editing model for real-world e-commerce content production. The model combines general image editing with e-commerce-specific capabilities, covering diverse e-commerce asset production needs such as adding/removing, replacing, composition adjustment, and local issue repair. It builds and maintains high-quality training samples through a multi-agent data engine, uses a training approach combining specialized capability optimization and multi-teacher distillation to integrate multiple capabilities, and has also built Ecom-Bench, an e-commerce-specific evaluation system with multiple tasks and dimensions. Relevant tests show that KwaiMind ranked first in overall score among participating open-source models on multiple general image-editing benchmarks and its self-built e-commerce evaluation; in an online A/B experiment for selecting product main image assets, it achieved a relative increase in actual
click-through rate of about 2.44%. Kuaishou’s technical team said it will continue to iterate and expand the model’s related capabilities.
@@@
TechWeb reports that at the online collective reception day for investors of listed companies in Chongqing jurisdiction and the semi-annual results briefing held on September 23, Xie Renjun, deputy general manager of Changan Automobile’s Capital Operations Department and deputy director of its Board Office, said the company’s solid-state battery project is advancing R&D in an orderly manner and has carried out validation of 20Ah oxide-polymer composite solid-state batteries. He also said Changan Automobile will accelerate the layout and industrialization of emerging industries such as embodied AI humanoid robots and flying cars, striving to mass-produce humanoid robots by 2028 and launch flying cars for commercial routes by 2030. Previously, Tan Huan, chief expert of Changan Automobile and general manager of Changan Tianshu Intelligent Robot Company, said at this month’s China Automotive Industry Development (TEDA) International Forum that Changan’s robotics layout will first
modularize automotive components such as cabins, central controls, and armrests, then move toward intelligent mobility ecosystem robots, covering handling, logistics, cleaning, energy storage, and energy replenishment, and further expand into extreme scenarios such as inspection, security, and firefighting. Its overall logic is that in the future, robots will build cars, robots will sell cars, and robots will be put into cars, ultimately achieving robots building robots and then robots building everything in the world.
@@@
According to QbitAI, Zheng Yinhe, head of AI NPC and Gameplay for miHoYo’s Honkai series, disclosed at the Apsara Conference that the company is advancing “AI entering games, and games feeding back into AI.” Earlier, co-founder Liu Wei, known as “Dawei Ge,” said the company would invest up to RMB 100 billion in AI over the next three years. In April this year, Honkai: Star Rail added AI dialogue for conductor Pom-Pom, generating more than 60 million conversations in a week, with one person chatting 1,379 times in a single day; AI Pom-Pom uses emotion tags to generate actions, strategy-enhanced RAG to maintain its persona, and short-, medium-, and long-term memory to remember player preferences. After being available for more than a month, it was temporarily taken offline for optimization, and needs to solve issues such as response stability, 42-day version updates, and concurrency costs for tens of millions of players. Zheng Yinhe also demonstrated an AI board game that lets AI
characters independently assess the situation and participate in gameplay, with dialogue able to affect relationships and plot. miHoYo’s internal EchoX platform uses multiple agents to collaborate on R&D, can read logs to troubleshoot performance, and generate playable demos from sketches. Jiang Dawei revealed that dozens of agents collaborated continuously for 13 hours, burning through RMB 2 million worth of tokens.
@@@
Alright, that’s all for today’s show. Thank you to every one of you for joining me. May you tuck gentleness into your pocket and carry good vibes forward. If you enjoy our show, we’d really appreciate it if you subscribe and share it with your friends. That helps us a lot. See you next time!
